MULTI-AGENT CONTENT PIPELINES
Separate agents draft, critique and fact-check, because one model reviewing itself always approves.
A single model asked to write and then check its own work will approve it every time. That is the whole reason content pipelines are multi-agent: the drafter, the critic and the fact-checker are separate runs with separate instructions, and the critic is given explicit permission to reject. Output quality comes from that separation far more than from a cleverer prompt.
Grounding is the second half. Drafts are written against a corpus of your own files — briefs, transcripts, product docs — searched with BM25 keyword retrieval, so a claim in the draft traces back to a line in a source document. Anything the corpus does not support gets flagged rather than smoothed over, which is how you avoid publishing a confident invented statistic.
I run a pipeline like this on my own publication, over a research corpus of roughly two hundred source files. It drafts, it gets rejected, it redrafts, and nothing reaches a live page without me reading it. That last part is not a limitation to be engineered away — a pipeline with no human gate publishes its mistakes faster than anyone can catch them.
WHAT I DO
- Role-split pipeline: drafter, critic and fact-checker as separate agents with separate instructions and a real reject path
- Brand voice as a checkable rules file — banned words, sentence-length limits, forbidden claims — kept in the repo, not implied in a prompt
- BM25 keyword retrieval over your own source files, with the matched lines attached to each draft for checking
- Claim flagging — anything the corpus cannot support is marked for a person instead of being published smoothly
- Publish step into your CMS or straight to Markdown files, gated on an explicit human approval
- Per-piece cost report: tokens and rupees for a draft, a critique round and a rewrite, so volume decisions use real numbers
HOW I WORK
- 01
CORPUS
I gather the source files the pipeline may draw from and get keyword search working over them before any writing happens.
- 02
RULES
Your voice becomes a checkable rules file — banned phrases, structure, claims that need a source — not a paragraph of adjectives.
- 03
PIPELINE
I wire drafter, critic and fact-checker with a reject path, then push real briefs through until the rejections are fair.
- 04
GATE
Approval and publishing go in last, so nothing can reach a live page before a person has read it.
BEST FOR
- A real backlog — dozens of pieces to write from source material that already exists
- Writing with a strict factual floor: documentation, research summaries, technical posts
- Teams whose drafts all read identically because one prompt is doing all the work
- Anyone who needs a draft traceable back to the brief or transcript it came from
NOT THIS
- Not content strategy. What to publish, for whom and in what order is positioning work that lives under brand solutions — this page is the machine that executes a plan you already have
- Not image or video generation. Visual production is a separate service; this pipeline handles words, structure and sources
- Not autopublishing. Every build ships with a human approval gate and I will not take it out on request
THE PROOF
Published case studies from ai orchestration work. Each one names the client and shows what was delivered.
STRAIGHT ANSWERS
REST OF AI ORCHESTRATION
MAKE A
WISH.
Tell me what you're building. I'll tell you how loud it can get.
- /01
YOU WRITE
Tell me what you're building and what's in the way. Two lines is enough.
- /02
WE TALK
A 30-minute call, booked within 7 days. Real questions, no pitch theatre.
- /03
YOU GET A PLAN
Scope, timeline and an honest number, usually within a week.